What is the difference between AI governance and AI security?
AI governance decides which AI systems an organisation uses, for what purpose, who owns them, what risk they carry and on what conditions they are approved, and it keeps the evidence. AI security protects those systems and their data at runtime, against leakage, misuse and attack. Governance is the decision and the record. Security is one of the ways the decision is enforced.
Neither replaces the tools you already run. DLP, CASB and SSE, DSPM, GRC platforms and LLM gateways each cover part of the AI problem. The useful question is not which one wins, but which question each one answers and what it cannot see.
How the categories compare
| Category | Question it answers | Unit of record | Where it acts | Typical owner | What it misses for AI |
|---|---|---|---|---|---|
| AI governance | Should we use this AI system, for this purpose, on these conditions? | The AI system, with owner, purpose, risk and approval | Decisions and evidence; enforced wherever it is connected to a control point | Risk, compliance, DPO, an AI governance lead | Without a runtime control, nothing stops traffic that ignores the decision |
| AI security | Is this request, model or agent safe right now? | The request, the model endpoint, the agent | Inline on AI traffic and on endpoints | CISO, security engineering | Why a system exists, who approved it and when that approval ends |
| DLP | Is sensitive data leaving through this channel? | The file, message or upload, matched to a pattern or label | Endpoint, email, network, cloud storage | Security operations, data protection | Data pasted or reworded into a prompt, and what the model may do with it |
| CASB and SSE | Which cloud services are used, by whom, and is access allowed? | The cloud application and the user session | Network edge and cloud APIs | Network security, security architecture | Often AI features inside approved SaaS; usually purpose, risk classification and approval history |
| DSPM | Where is sensitive data stored and who can reach it? | The data store and the data asset | Cloud and on-premises data stores | Data security, cloud security | Data in transit to a model, and the AI system that consumes it |
| GRC | Are our risks and controls documented, owned and tested? | The risk, the control, the policy, the audit finding | Workflows and registers, not traffic | Risk management, internal audit, compliance | Live facts about AI use; records drift unless something feeds them |
| LLM gateway | Which model serves this request, at what cost, under what limits? | The API request and response | Between applications and model providers | Platform or AI engineering | Ownership, purpose and approval, unless tied to a governance record |
Why prompts are hard for classic file DLP
Classic DLP was designed around files and messages that move as a whole. It looks for a pattern, a fingerprint of a known document or a sensitivity label, and acts when the object crosses a boundary. A prompt breaks most of those assumptions.
- No file, no label. Text pasted from a labelled document into a chat box carries none of the label. The fingerprint of the source document no longer matches.
- The data is rewritten. People summarise, translate or paraphrase before they paste. A contract clause rephrased in plain words matches no pattern.
- The channel looks ordinary. The request is an HTTPS call to a well-known domain, often from a browser session with a personal account.
- The response matters too. A model can return data, code or instructions. File DLP does not ask what came back or what an agent did with it.
- Agents act. An agent connected to tools can read and change systems. That is an authorisation question, not a content one.
Structured identifiers such as card numbers, national identifiers and IBANs can still be caught in a prompt by pattern rules, if something inspects the request before it leaves. Unstructured business context usually cannot. That is why AI controls combine content inspection with decisions about which systems may receive which categories of data at all.
Where they overlap
- Discovery. CASB, SSE, proxy logs and an AI gateway can all show which AI services are contacted. The governance record turns those sightings into owned systems.
- Data rules. DLP, SSE and an AI gateway can all inspect content. The governance record says which data categories each approved system may receive.
- Risk. A GRC platform holds the enterprise risk register. AI governance supplies the per-system classification and evidence that feed it.
- Logging. Every layer produces logs. Only some of them tie a request to the system, owner and approval it belongs to.
How they fit together
| Layer | Role | Typical tools |
|---|---|---|
| Enterprise risk and policy | Sets risk appetite, policies and the control framework | GRC platform, policy library |
| AI governance record | Per-system owner, purpose, risk, reviews, approval and evidence | AI governance platform or a maintained register |
| AI admission and runtime | Applies the approval to each request: models, data, budgets, agents | AI gateway, AI security controls |
| Network and cloud access | Sees and controls access to cloud services, including AI domains | CASB, SSE, secure web gateway |
| Data protection | Finds and protects sensitive data where it is stored and where it moves | DSPM, DLP, classification |
Read the stack from the top. Policy says what is acceptable. The governance record applies it to each AI system. The admission layer enforces it on each request. Network and data controls catch what never passed through the admission layer.
Where to start: a decision guide
| If your problem is | Start with |
|---|---|
| Nobody knows which AI systems are in use | An AI inventory fed by proxy and DNS logs, a gateway and declarations |
| Employees paste customer data into public chat services | A sanctioned alternative routed through a gateway with prompt inspection, plus CASB or SSE rules for the rest |
| Sensitive data sits in stores that AI pipelines read | DSPM to map the data, then governance records for the systems that consume it |
| Auditors or regulators ask how AI use is approved | An AI governance record with reviews, approvals and an evidence trail |
| Agents can call tools that change production systems | An agent and tool inventory with risk review, and identity checks at the gate |
| AI spend is growing without an owner | A gateway with team budgets and a cost ledger tied to each system |
| Enterprise risk reporting lacks AI | Feed per-system AI classifications into the existing GRC register |
A practical path
- Map what you own. List which of DLP, CASB or SSE, DSPM, GRC and gateway you run, and what each inspects on AI traffic today.
- Build the inventory. Record each AI system with an owner, a purpose and the data categories that can reach it.
- Classify and decide. Assess risk from the facts, and approve with conditions and an end date.
- Choose one enforcement point. Route sanctioned AI traffic through a gate that applies the approval to each request.
- Close the gaps with existing tools. Use CASB, SSE and DLP for traffic that does not pass the gate.
- Feed the risk register. Send AI classifications and evidence into GRC so enterprise reporting includes them.
Checklist
- Each AI system has an owner, a purpose and a risk classification
- Each approval has conditions and an end date
- You know what your DLP and SSE inspect on AI domains
- Sanctioned AI traffic passes one point that can refuse a request
- Agents and their tools are recorded and reviewed
- AI risks appear in the enterprise risk register
- Evidence can be produced per system without rebuilding it by hand
Sources: OWASP Top 10 for LLM Applications 2025, which lists prompt injection, sensitive information disclosure and excessive agency among the main risks for LLM applications.
How MERIDVAR fits
MERIDVAR combines an AI governance record and an admission gateway that can refuse requests, in one self-hosted, single-tenant appliance that you install and operate on your own network. Nothing is sent to MERIDVAR.
- The record. An AI inventory with owner, purpose, vendor, models and provenance, rules-based risk classification, privacy and security reviews, and approvals by an administrator who did not make the request.
- The gate. An admission gateway applies approved model lists, prompt DLP for structured identifiers, data residency, team budgets and agent signatures. A refused request never reaches the model provider.
- The evidence. An append-only, hash-chained evidence log and an audit package per AI system or vendor.
It does not replace DSPM, CASB, SSE or GRC. Controls apply only to traffic routed through the gateway, and the product is in beta. See AI governance and the product overview.